Congestion Control

Congestion control is the set of rules by which senders and the network decide how much data may be in flight at once.

9 frameworks, from 1988 to today · Networks · Computer Science

still a working view ended · each leads to its entry below

Overview

Congestion control is the set of rules by which senders and the network decide how much data may be in flight at once1. When many computers share a link, a switch, or a radio channel, demand can exceed capacity1. Packets then queue, delay rises, and buffers overflow1. Congestion control is what keeps that from collapsing into a state where everyone sends as fast as possible and almost nothing gets through1. It is one of the load-bearing ideas of the Internet: the network itself is deliberately simple, and most of the intelligence that keeps it stable sits in the endpoints. The history of the field is the history of how that endpoint intelligence was invented, refined, and repeatedly rethought as the network changed around it.

The collapse of the early Internet and the first fix

In the mid-1980s the Internet suffered a series of "congestion collapses." Throughput on some paths fell by orders of magnitude while hosts kept retransmitting1. The cause was that the early Transmission Control Protocol (TCP) had no notion of how much the network could absorb. A sender pushed data until a packet was lost, then retransmitted, and the retransmissions themselves added load, making loss more likely1. The network had no feedback loop to tell senders to slow down.

The fix came from Van Jacobson, whose 1988 paper with Michael Karels described the algorithms that became TCP's congestion control. The central idea was the congestion window: a limit, maintained by the sender, on how many unacknowledged bytes may be outstanding2. The sender grows the window when acknowledgments arrive and shrinks it when packets are lost2. Jacobson's scheme had two phases. Slow start begins with a small window and doubles it each round trip, so a connection ramps up quickly from nothing. Congestion avoidance then grows the window by roughly one packet per round trip, a cautious probing of spare capacity. On loss, the sender cuts the window, classically in half2.

This is additive increase, multiplicative decrease (AIMD)3. The name describes the shape of the control: the window climbs linearly while things go well and drops sharply when they don't. The reason it works is a property of how competing senders share a bottleneck. If several AIMD flows share one link, their windows converge toward a fair split, because a flow that is getting more than its share loses more when congestion hits3. The scheme is loss-based: it treats a dropped packet as the signal that the network is full. Jacobson also added fast retransmit and fast recovery, so a sender could react to a few duplicate acknowledgments without waiting for a timeout. These algorithms, refined through the late 1980s and standardized in the early 1990s, are the foundation on which nearly everything since builds.

The router's role: active queue management

Loss-based control has a weakness. If a router's buffer is large, it fills completely before any packet is dropped, so every sender learns about congestion only after long queues have built up4. The delay is felt by everyone sharing the link. The alternative was to have routers signal congestion earlier, before buffers overflow.

The first widely discussed proposal was the DECbit scheme from the late 1980s, in which routers set a congestion bit in packets when average queue length crossed a threshold. The idea reached the Internet community through Random Early Detection (RED), published by Sally Floyd and Van Jacobson in 1993. RED routers monitor the average queue length and, as it rises toward a maximum threshold, drop arriving packets with a probability that increases with the queue1. The randomness matters: it spreads the signal across many flows so that they back off together rather than in lockstep. This family of techniques is called active queue management (AQM): the router manages its queue actively rather than letting it fill and overflow passively15.

RED was influential but hard to tune, and deployment was limited. Its descendants and rivals—among them variations that drop based on the queue's instantaneous length or that target a fixed delay—kept the idea alive. AQM matters because it moves part of the congestion signal from the endpoint's inference ("a packet was lost, so somewhere is full") to the network's explicit or statistical hint. That shift becomes central to later developments.

Explicit signals and rate-based control

A separate line of work asked whether the network could tell senders directly what rate to use, rather than making them infer it from loss. Explicit Congestion Notification (ECN), specified in the late 1990s and early 2000s, lets a router mark a packet instead of dropping it, and the receiver echoes the mark back so the sender slows down without any loss6. ECN is a modest change but it removes the ambiguity between congestion loss and other kinds of loss.

More ambitious was explicit rate-based control, in which routers compute and communicate a sending rate. The Available Bit Rate (ABR) service in ATM networks, developed in the 1990s, had switches feed back explicit rates to sources. In the Internet, the Rate Control Protocol (RCP) proposed that routers calculate a fair rate for each flow and stamp it into packets, so senders need not probe at all. These designs promised fast convergence and low queues, but they required routers to keep per-flow state, which conflicted with the Internet's preference for stateless, simple forwarding. They remained largely research proposals, though their ideas about explicit feedback resurface in modern data-center designs.

When loss stops meaning congestion

By the late 1990s, wireless links exposed a flaw in loss-based control. On a radio link, packets are lost to interference, fading, and handoffs, not because the network is congested. A TCP sender cannot tell the difference, so it halves its window on every such loss and performs badly. The field's response took several forms.

One approach was to hide wireless loss from the sender: split the connection at the wireless edge, or add link-layer retransmissions so the radio's losses are repaired locally. Another was to make the sender smarter. TCP Westwood, from the early 2000s, estimated the bottleneck bandwidth from the rate of returning acknowledgments and set the window from that estimate after a loss, rather than blindly halving. A third line used delay as the congestion signal instead of loss. If a sender watches the round-trip time, rising delay indicates that queues are building—congestion—while a lost packet on a lossy link may say nothing about congestion. Delay-based control had been explored as early as the 1980s, and it returned in schemes such as TCP Vegas (mid-1990s), which compared the measured and expected round-trip times to keep queues small. Vegas was fair and gentle but lost out in competition with loss-based flows, which grab more bandwidth when sharing a link. This fairness problem—delay-based flows being crowded out by aggressive loss-based ones—became a recurring theme.

Model-based estimation

A more systematic answer to "what is the network's capacity?" came from model-based bandwidth and delay estimation. Instead of reacting to individual loss events, a sender builds a model of the path and uses it to set its rate. The best-known example is TCP BBR, published by Neal Cardwell and colleagues in 2016. BBR periodically measures the bottleneck bandwidth (from the rate at which acknowledgments return) and the minimum round-trip time (the path's propagation delay with empty queues). It then sends at a rate matched to that model, probing upward occasionally to check whether more capacity has appeared. Because it does not treat loss as the primary congestion signal, BBR performs well on lossy links and keeps queues short on many paths. It also raised new questions: BBR flows can be unfair to loss-based flows sharing a bottleneck, and its behavior in the presence of deep buffers and competing traffic is still studied.

Congestion control beyond TCP

TCP was never the only transport. The User Datagram Protocol (UDP) carries no congestion control at all, and applications built on it—streaming media, voice, games—had to add their own or risk starving the network. The Datagram Congestion Control Protocol (DCCP), standardized in the mid-2000s, offered a framework for congestion-controlled datagram traffic, with pluggable algorithms. Meanwhile, the rise of real-time media drove work on rate-based control for video and voice, where the goal is a smooth sending rate rather than a window that jumps. The Real-time Transport Protocol (RTP) and its control protocol carried feedback that let senders adapt. These efforts established that congestion control is a general problem for any sender, not a TCP-specific one, and that different applications want different trade-offs between smoothness, latency, and fairness.

The data center rewrites the rules

The 2010s brought a setting where the old assumptions broke down: the data center. Here the network is owned by one operator, round-trip times are microseconds, links are fast, and switches have shallow buffers. Loss-based TCP reacts too slowly and its timeouts are far too coarse. The result was a burst of new designs.

Data Center TCP (DCTCP), from Mohammad Alizadeh and colleagues in 2010, used ECN marks to let senders react to the fraction of marked packets, keeping queues very short. Priority Flow Control and lossless fabrics addressed buffer overflow at the link layer. A wave of schemes followed—TIMELY used round-trip-time gradients; DCQCN brought congestion control to remote direct memory access (RDMA) traffic; Swift combined delay and loss signals. The common thread is low-latency congestion control: the goal is not just high throughput but keeping queues near empty, because in a data center a millisecond of queuing is a large fraction of the total delay. These designs also leaned on programmable switches and software-defined networking, which made it practical to give the network a more active role in signaling.

Where the field stands

Today congestion control is a family of approaches rather than a single algorithm. Loss-based AIMD remains the baseline, embedded in TCP and in the congestion control of QUIC, the transport that now carries much web traffic. Delay-based and model-based schemes (Vegas descendants, BBR) compete with it, and hybrid designs blend loss, delay, and explicit signals. AQM is enjoying a revival, with algorithms designed to keep queuing delay low for latency-sensitive traffic. Data-center and low-latency designs have pushed the field toward explicit feedback and fine-grained control, while wireless and satellite networks keep the lossy-link problem alive.

The open questions are old ones in new clothes. How should flows that use different signals share a link fairly? How should congestion control behave when the path crosses technologies with wildly different characteristics? Can machine learning set rates better than hand-designed rules, and can such controllers be trusted? And as the network becomes more programmable, how much of the control should move from the endpoint into the network itself? The answers will be written by the same kind of reasoning that produced the congestion window: measure what the network is doing, model it, and adjust.

References

  1. Network congestion (Congestion control, Network capacity, introduction, Congestive collapse, Random early detection, Active queue management), Wikipedia
  2. TCP congestion control (Congestion window, TCP Tahoe and Reno), Wikipedia
  3. Additive increase/multiplicative decrease, Wikipedia
  4. Random early detection, Wikipedia
  5. Active queue management, Wikipedia
  6. Explicit Congestion Notification, Wikipedia

The frameworks of Congestion Control

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1988-Present

    Additive increase/multiplicative decrease

    How should senders adjust their rate to share a network fairly?

    Fly to its tower in Nousfarer

  2. 1993-Present

    Active queue management

    When should routers signal congestion before buffers overflow?

    Fly to its tower in Nousfarer

  3. 1994-Present

    Delay-based congestion control

    Can rising delay tell a sender the network is congested?

    About Delay-based congestion control
    In the 1990s, as the Internet grew beyond its research origins, congestion control faced a fundamental question: could a sender detect incipient congestion before packets were actually dropped? Delay-Based Congestion Control offered a distinctive answer. Instead of waiting for the blunt signal of packet loss, this framework proposed that a sender should monitor changes in round-trip time (RTT)—the time a packet takes to travel to the receiver and back. Rising delays, the argument went, reveal growing queue occupancy inside routers, giving an early warning that the path is becoming congested. This made delay a gentler, more proactive signal than loss, and it opened up a new design space for transport protocols that could react before buffers overflowed.
    Delay-Based Congestion Control is a member of the End-to-End family—it keeps the network core simple and places all intelligence at the endpoints—but it diverges sharply from the dominant AIMD/TCP paradigm. Where AIMD/TCP treats the network as a black box that only reveals congestion through dropped packets, delay-based schemes treat the network as a transparent pipe whose internal state can be inferred from timing. The core curriculum for this framework centers on a handful of concepts that define how delay is measured, interpreted, and turned into a control signal.
    Core Curriculum
    1. Round-Trip Time (RTT): The fundamental observable. The sender measures the time between sending a packet and receiving its acknowledgment. This is the raw material from which congestion signals are extracted.
    1. Base RTT: The minimum observed RTT when the path is unloaded. It serves as a reference point: any increase above base RTT is attributed to queuing delay in routers.
    1. Queuing Delay: The component of RTT that varies with congestion. It is the difference between current RTT and base RTT, representing time spent waiting in router buffers.
    1. Self-Clocking: The insight that acknowledgments returning from the receiver act as a natural clock for the sender. Each ACK signals that a packet has been delivered, allowing the sender to pace new transmissions without needing an external timer.
    1. Window-Based Control: The sender limits how many packets it can have in flight at once, expressed as a congestion window. Delay-based schemes adjust this window in response to RTT changes.
    1. Rate-Based Control: An alternative to window-based control where the sender regulates its transmission rate (packets per second) directly. Some delay-based protocols use rate rather than window to match the observed path capacity.
    1. Additive Increase / Multiplicative Decrease (AIMD): The classic rule borrowed from TCP: gradually increase the window when the path seems uncongested, and cut it sharply when congestion is detected. Delay-based frameworks adapt this rule to use delay thresholds rather than loss events.
    1. Delay Threshold: A predefined RTT increase (e.g., a doubling of base RTT) that triggers a multiplicative decrease. This replaces the loss-based trigger of AIMD/TCP.
    1. Queue Occupancy Estimation: The sender estimates how many packets are queued ahead of it by comparing current RTT to base RTT and factoring in the bottleneck link rate. This turns delay into a proxy for queue depth.
    1. Fairness: The property that multiple competing flows should converge to roughly equal bandwidth shares. Delay-based frameworks must ensure that their control rules do not let one flow starve others.
    1. Stability: The requirement that the sender's window or rate should not oscillate wildly. Delay-based schemes must avoid overreacting to transient RTT spikes.
    1. Burstiness: The tendency of a sender to transmit packets in clusters rather than a smooth stream. Delay-based control can reduce burstiness by pacing transmissions with ACK arrivals.
    Intellectual History
    Delay-Based Congestion Control emerged in the mid-1990s as a direct response to the limitations of AIMD/TCP. Van Jacobson's 1988 TCP congestion control algorithms had saved the Internet from congestion collapse, but they relied entirely on packet loss as a signal. Loss is a late indicator—it only appears after a buffer has already overflowed. Researchers began asking whether a sender could detect the buildup before the drop.
    The seminal work came from Raj Jain in 1994, who proposed a scheme called TCP Vegas. Jain showed that a sender could estimate the queue occupancy at the bottleneck router by comparing current RTT to the minimum observed RTT (base RTT). When the estimated queue grew beyond a threshold, the sender would reduce its window. Vegas was the first complete delay-based congestion control protocol, and it demonstrated that delay could serve as a reliable, early congestion signal. Its key insight—that queuing delay reveals incipient congestion—became the foundation of the entire framework.
    TCP Vegas was never widely deployed, but it inspired a line of research that continued through the late 1990s. TCP Reno (1996) incorporated some of Vegas's ideas, though it remained primarily loss-based. TCP NewReno (1998) improved loss recovery but did not adopt delay-based control. The framework's influence was more conceptual than practical during its dominance window: it established that delay was a viable signal, and it set the stage for later protocols that would combine delay with other indicators.
    Delay-Based Congestion Control coexisted with its peer frameworks throughout the 1990s. Router-Assisted schemes (DECbit, RED, ECN) asked routers to mark packets before dropping them, but they required router changes that were slow to deploy. Equation-Based Congestion Control (TFRC, 2000) would later model the path's behavior mathematically, using delay as one input among many. High-Speed Congestion Control (2002) addressed the problem of filling large pipes without loss, and it often used delay-based pacing. Hybrid frameworks (2005) combined loss, delay, and ECN signals. Model-Based Congestion Control (BBR, 2016) would eventually use delay to estimate the bottleneck bandwidth, but it did so through a model of the path rather than direct queuing delay.
    Delay-Based Congestion Control never replaced AIMD/TCP as the default paradigm, but it permanently changed the conversation. It showed that the network's internal state was readable from the endpoints, and it gave later frameworks a vocabulary for talking about queuing delay, base RTT, and self-clocking. Its legacy is not a single dominant protocol but a set of concepts that every congestion control designer must now understand.
    The 8 concepts its tower teaches
    • Additive Increase / Multiplicative Decrease (AIMD)
    • Base RTT
    • Delay Threshold
    • Queuing Delay
    • Rate-Based Control
    • Round-Trip Time (RTT)
    • Self-Clocking
    • Window-Based Control

    Fly to its tower in Nousfarer

  4. 1994-Present

    Explicit congestion signaling

    Should the network tell senders directly how fast to send?

    Fly to its tower in Nousfarer

  5. 1996-Present

    Datagram and real-time media congestion control

    How should real-time and datagram traffic control its sending rate?

    Fly to its tower in Nousfarer

  6. 2002-Present

    Explicit Rate-Based Congestion Control

    Can routers compute a fair rate for each flow?

    Fly to its tower in Nousfarer

  7. 2010-Present

    Low-latency data-center congestion control

    How can data-center traffic keep latency very low?

    Fly to its tower in Nousfarer

  8. 2016-Present

    Model-based bandwidth and delay estimation

    Can a sender model the path's capacity and delay?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Additive increase/multiplicative decr…Active queue managementDelay-based congestion controlExplicit congestion signalingDatagram and real-time media congesti…Wireless and lossy-link congestion co…Explicit Rate-Based Congestion ControlLow-latency data-center congestion co…Model-based bandwidth and delay estim…

Beyond the field

Where Congestion Control's ideas came from in other fields, and what they led to there.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Congestion Control is a planet in Nousfarer

Its city has a tower for each of these 9 frameworks. Land, walk in, and climb one concept at a time.

Land on Congestion Control

Related fields

Fields its ideas are linked to

More in Networks

See where Congestion Control sits on the Map of Human Knowledge